China AI Native Industry Insights – 20250730 – Zhipu AI | Alibaba | Alibaba | more
Explore Zhipu AI’s GLM-4.5 model and Alibaba’s latest AI innovations, including the Qwen3-30B-A3B-Instruct-2507 model and open-source video models with cinematic aesthetics. Discover more in Today’s China AI Native Industry Insights.
1. Zhipu AI Launches GLM-4.5: Leading Open-Source Model for Agents and Code
🔑 Key Details:
– GLM-4.5 is a new flagship model for agent applications, available on Hugging Face and ModelScope under the MIT License.
– It incorporates a mixture of experts (MoE) architecture with 355 billion parameters, delivering state-of-the-art capabilities in reasoning, coding, and agent tasks.
– Fast and low-cost API at 0.8 CNY per million tokens input and 2 CNY output, with a speed of up to 100 tokens/sec.
– Achieved top performance in 12 benchmarks, ranking third globally for all models and first among open-source models.
💡 How It Helps:
– AI Developers: Access to a high-performing, open-source model encourages innovation in developing complex applications.
– Enterprises: Cost-effective deployment options facilitate wide-scale integration of advanced AI capabilities.
🌟 Why It Matters:
GLM-4.5’s launch signifies a crucial advancement in the open-source AI landscape, particularly in enhancing the capabilities for software development and intelligent agent applications. By offering high performance with lower costs, it challenges existing models and sets new standards for efficiency and functionality in AGI tools.
Original Chinese article: https://mp.weixin.qq.com/s/Psb5TJSFszReCQ8SwnjFyA
English translation via free online service: https://translate.google.com/translate?hl=en&sl=zh-CN&tl=en&u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%2FPsb5TJSFszReCQ8SwnjFyA
Video Credit: The original article
2. Alibaba Launches Qwen3-30B-A3B-Instruct-2507: New AI Model Unveiled
🔑 Key Details:
– Qwen3‑30B‑A3B‑Instruct‑2507 is an updated non‑thinking instruction model; it never emits and no enable_thinking flag is needed.
– MoE design: 30.5B params (128 experts, 8 active; ~3.3B activated), 48 layers, GQA 32Q/4KV; 29.9B non‑embedding.
– Native 256K context (262,144) for long documents and multi‑turn workflows.
– Capability gains: stronger instruction following, reasoning, math, science, coding, tool use; better long‑tail and multilingual coverage; improved alignment in open‑ended tasks.
– Tooling: on Hugging Face / ModelScope; supported in transformers ≥4.51.0; serve via sglang ≥0.4.6.post1 or vLLM ≥0.8.5; local runtimes include Ollama, LM Studio, MLX‑LM, llama.cpp, KTransformers.
💡 How It Helps:
– AI developers: 256K context plus strong instruction tuning boosts coding, tool use, retrieval, and long‑form tasks; OpenAI‑compatible serving and MCP tool integrations speed agent builds.
– Enterprises: deterministic non‑thinking outputs reduce latency and variance; multilingual coverage supports global teams; familiar stacks simplify deployment and scale.
🌟 Why It Matters:
Qwen3‑30B‑A3B‑Instruct‑2507 advances a practical non‑thinking, long‑context path for agents and coding assistants. Its MoE efficiency, broad ecosystem support, and solid benchmarks enable faster, more stable applications that handle long documents and complex toolchains at scale.
Original Chinese article: https://modelscope.cn/models/Qwen/Qwen3-30B-A3B-Instruct-2507
English translation via free online service: https://modelscope-cn.translate.goog/models/Qwen/Qwen3-30B-A3B-Instruct-2507?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en-US&_x_tr_pto=wapp
Video Credit: Qwen (@Alibaba_Qwen on X)
3. Alibaba Launches Open-Source Video Models with Cinematic Aesthetics
🔑 Key Details:
– Open-source video generation model ‘Wan2.2’ launched by Alibaba.
– Features three models: Wan2.2-T2V-A14B, Wan2.2-I2V-A14B, and Wan2.2-IT2V-5B, utilizing a total of 27B parameters.
– Introduces the first industry-controlled cinematic aesthetics system for advanced visual qualities.
💡 How It Helps:
– AI Developers: Access to innovative video generation models for enhanced creativity and resource-efficiency.
– Content Creators: Ability to produce high-quality video content with refined cinematic control, improving project outcomes.
🌟 Why It Matters:
This launch reinforces Alibaba’s position as a leader in AI-driven video generation, highlighting advancements in computational efficiency and artistic control. It addresses growing demands for sophisticated video tools, shaping new standards within the content creation industry.
Original Chinese article: https://mp.weixin.qq.com/s/iPL7OLQhwYdoFelHt41N6Q
English translation via free online service: https://mp-weixin-qq-com.translate.goog/s/iPL7OLQhwYdoFelHt41N6Q?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en-US&_x_tr_pto=wapp
Video Credit: The original article
That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.